These examples show how Epistamate behaved in controlled research and workflow experiments. They are not client case studies, comparative accuracy benchmarks or claims of universal superiority.
A specialist advisory-method experiment tested what happens when new client material arrives after propositions, diagnostics, findings and recommendations already exist. The objective was not to replace the advisory method, but to see whether a governed workspace could preserve the causal chain between new evidence and affected work.
New material was represented locally and a priority review set was surfaced without discarding the lower-priority source content.
Uncertainty could become a clarification question; client answers returned as attributed evidence rather than informal memory.
Accepted evidence could affect proposition → diagnostic → finding → recommendation → deliverable in order, rather than rewriting the end state directly.
A supplied industry fraud/cyber paper contained 27 citations. Epistamate checked route reachability separately from citation identity, claim support and substantive validity instead of treating a working bibliography link as proof that the nearby claim was established.
A live URL could still resolve to the wrong edition or fail to support the nearby claim. Those questions remained separate review states.
Preserved source contents were de-duplicated before later analysis so citation volume did not masquerade as evidentiary independence.
On AI coding-assistant productivity, 14 planned targets produced 12 acquired sources while 11 knowledge gaps and 5 validation leads survived the first synthesis into the next iteration.
The policy/public-sector experiment assessed the concluded EU–India FTA for German/EU manufacturers considering India expansion over a 3–5 year horizon, focused on automotive and pharmaceuticals. The framing explicitly separated negotiated FTA effects from India’s broader market growth, industrial policy, regulation, FDI, logistics and state-level operating conditions.
Tariff gains could not be read independently of quotas, phase-ins, classification and origin eligibility.
Potential tariff savings coexisted with approval, pricing, regulatory and product-status constraints.
The second iteration materially added rules-of-origin detail, evidence on zero uptake of the SPMEPCI scheme, pharma FDI material and firm-level manufacturing evidence while retaining major gaps.
Iteration 2 represented 2,168 passages. The first bounded pass analysed 67. The researcher then requested two non-overlapping 60-passage expansions, taking the admitted claim state from 27 to 43 and then 61. Later evidence mostly deepened an increasingly stable research position rather than creating a completeness claim.
A regulatory/compliance experiment examined supplier terms from a leading IT-services provider against the surrounding legal and standards context. The central test was whether Epistamate could keep contractual assertions separate from binding law, official guidance, standards material and unresolved engagement facts.
Binding law, official guidance, standards and contract language could all matter without being flattened into one confidence hierarchy.
A regulatory source could be authoritative and still require jurisdiction, role, service configuration or contractual facts before a conclusion applied.
Unknown engagement facts stayed as explicit diligence needs rather than being filled by inference.
These examples describe controlled product behaviour and research state. They do not constitute legal advice, policy recommendations, audit opinions, client endorsements or quantified performance claims.